Application of an Empirical FBRM Model to Estimate Crystal Size

Jan 21, 2014 - fixing the CSD as a system input and the chord-length distribution (CLD) from the FBRM as the output, we constructed the model as a lin...
1 downloads 0 Views 2MB Size
Article pubs.acs.org/crystal

Application of an Empirical FBRM Model to Estimate Crystal Size Distributions in Batch Crystallization Huayu Li, Yoshiaki Kawajiri,* Martha A. Grover, and Ronald W. Rousseau School of Chemical and Biomolecular Engineering, Georgia Institute of Technology, Atlanta, GA 30332, United States ABSTRACT: Monitoring crystal size distributions in situ is a challenge in crystallization engineering. Such a capability provides comprehensive information on the crystallization, which can be used to improve the quality of products. We developed an empirical focused-beam reflectance measurement (FBRM) model to provide such estimates. By experimentally fixing the CSD as a system input and the chord-length distribution (CLD) from the FBRM as the output, we constructed the model as a linear transformation from CSD to CLD. A regularized least-squares algorithm, which considers the mass balances on solid and liquid phases, was used to estimate the CSD. The results of batch cooling crystallization experiments show that estimates of the evolution of CSD are in accordance with the general understanding of crystallization kinetics, and the CSD estimates of final product agreed with the CSD obtained from sieving. The simplicity and practicality of the model make it a significant enhancement to the use of FBRM in monitoring, simulating, and controlling crystallization processes.



temperature to dissolve fine particles so that a desired crystal size distribution (CSD) is obtained.2,18−20 To extract quantitative information from a CLD, several models of the FBRM have been developed to explain and predict the mapping between a PSD and a CLD. With a reliable model, the PSD can be estimated in real time and kinetic models can be constructed. First-principles models for FBRM have been established to convert a CLD to a CSD from the physical description of how a chord is measured. In the geometric model, it is assumed that a chord is a straight line across the 2D projection of a particle. Therefore, the model relies on the geometry of the crystals and simulates the 2D projection of the crystals. The CLD can be calculated by convoluting the probability of measuring certain lengths from the projection of a given particle onto the probability of getting the size from a known PSD.21,22 Li et al.23 have found that for opaque particles (e.g., ceramic beads, zinc dust, and plasma aluminum), the geometric model can simulate the CLD and thus the PSD can be well-estimated by constrained leastsquares methods. However, when applied to crystallization processes, it has been reported that the geometric model is unable to estimate the CLD from a given CSD.24−26 Because crystals have transparent facets, the laser may be reflected or refracted and light scattering is not uniform across their surfaces. In addition, the backscattered laser intensity also depends on the distance between crystals and the probe. These two factors lead to a phenomenon called “chord splitting,” in which a long scanning

INTRODUCTION Crystallization is a key separation and purification technique in the pharmaceutical and fine chemical industries, and many online tools have been developed to monitor and control crystallization processes. Focused beam reflectance measurement (FBRM) is one of the in situ sizing tools, which collect valuable information on the particle size distribution. It employs a rotating laser beam to scan the fluid medium and records the back-reflected laser. An intensity sensor measures the lengths of the scanned paths across particles and defines those lengths as chords. In this manner, the chord length distribution (CLD), which is a measure of the particle size distribution (PSD), is obtained in situ and in real time. This makes FBRM a very promising tool to analyze and control particulate processes, such as crystallization,1,2 precipitation,3,4 granulation,5,6 and emulsion.7,8 The relationship between the CLD and the PSD has been the subject of many investigations. Monnier et al.9 first showed that although CLD measurements are affected by many other settings, such as particle properties, focal point of the laser beam, positioning of the probe, and agitation rate, the total chord count is positively correlated with total crystal concentration in several solvents, and the mean size of the CLD also grows with the mean size of particles. Similar conclusions have been reported for other suspensions.10−12 This qualitative relationship has been used to monitor nucleation and growth behaviors in crystallization: for example, determining metastable zone width1,13,14 and investigating crystallization kinetics.15−17 Moreover, FBRM can be used as a qualitative indicator to detect nucleation; upon the moment of nucleation, control actions can be designed to raise the © 2014 American Chemical Society

Received: October 7, 2013 Revised: December 20, 2013 Published: January 21, 2014 607

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

length density function for a single crystal. The CLD density function qL0(s) also depends on the FBRM settings (for instance, the sampling time) but all those factors are incorporated in qp(s,L0) as long as the settings are unchanged. The CLD density function q(s) is obtained by integrating the contribution from every crystal size

path on a crystal surface is detected as multiple short chords due to the fluctuation of the backscattered light intensity. Therefore, an optical model considering these effects of crystallization systems has been built27 and shown to be able to predict the CLD of lactose crystals.28 Nevertheless, the optical model is complex and requires many process parameters, including refractive indices of the solid and liquid, the shape of the crystals, and the laser intensity profile; it also requires knowledge of FBRM algorithms on how to define a chord length from the reflectance intensity. The model contains nonlinear expressions and usually uses Monte Carlo simulations. Such a nonlinear model may not be suitable for online monitoring unless it can be simplified. Additionally, it is difficult to collect comprehensive data for every property of the system to construct the computation-intensive model. In our previous study,25 we proposed and validated an empirical approach to modeling FBRM measurements in a crystal−nonsolvent suspension. The model is considered as a gray box and analyzed by studying the input−output responses under certain principles and hypotheses, in which the input is the CSD and the output is the CLD. Assuming that some factors, such as the crystal shape, optical properties, operating conditions, etc. are statistically constant during the crystallization process, we mainly focused on the effect of the number of crystals at different sizes (i.e., the CSD is classified based on one size dimension). Therefore, the model is a linear CSD-toCLD transformation matrix, constructed from the characteristic CLD at each crystal size. The measurement of the characteristic CLD, or fingerprint CLD as we call it, was accomplished by sieving the crystals into different size ranges and sequentially adding them into a nonsolvent of the crystals. By observing the change of CLD, the fingerprint CLD can be obtained for each size range. In the present work, an FBRM model was built for batch cooling crystallization of paracetamol from ethanolic solution. Crystals were sieved into nine size ranges to perform the CLD fingerprint measurements in a saturated solution. Both the FBRM and an attenuated total reflectance-Fourier transform infrared (ATR-FTIR) probe, which measures the solute concentration, were applied to batch cooling crystallization. An algorithm unifying CLD and concentration was developed to estimate the CSD. Experimental results demonstrate that this approach is able to track the change of CSD during crystallization.





qp(s , L)n(L) dL

(2)

In practice, the CLD from the FBRM is a histogram, and the CSD is also a histogram since it is obtained from sieving. Thus, we rewrite eq 2 as a discretized form by defining

bi =

si + 1

∫s

q(s) ds ,

i = 1, 2, ..., p

i

xj =

∫L

Lj + 1

n(L) dL ,

j = 1, 2, .., r

j

(3)

(4)

where q and n are discretized in to p and r bins. Equation 2 can be reformulated into a discrete form: (5)

b = Ax

where b ∈ p and x ∈ r are CLD and CSD histograms, respectively. If the single crystal CLD density function qp(s,L) for crystals in a size bin [Lj, Lj+1] is assumed to be constant and the geometric mean of LjLj+1 is taken as the representative size for the jth bin, each element in the matrix A, can be written as

∫s

Ai , j =

si + 1

i

qp(s ,

LjLj + 1 ) ds

(6)

In this work, we determine A experimentally instead of a firstprinciples model because we do not have the complete information on how a chord is generated and what factors have an impact on the measurement. From a statistical point of view, we assume that there is a fingerprint CLD histogram ai for a given crystal size range (i.e., these are the column vectors of A), and eq 5 can be rewritten as the following equation. r

b=

∑ xiai i=1

(7)

Note that the fingerprint ai is an intensive property, which is independent of the number of the crystals in the suspension. The fingerprint for size range i is obtained in the following procedure: (1) the crystals in size range i are obtained from sieving, and these crystals are added to a saturated solution at multiple times. For each addition, the mass of crystals being added is measured, in order to calculate the crystal concentration in the suspension [no. of crystals/milliliter]. (2) After the jth addition, the crystal concentration xj was calculated. Since only xi,j is nonzero and xk,j = 0 for k ≠ i, xj can be calculated by the following: mcrystal xi , j = kV ρli3Vsolution (8)

METHODS AND MATERIALS

Model Development. Crystal size distribution (CSD) is a description of a categorization of crystals based on their sizes, which can be either a histogram or a density function. The density function precisely depicts how the crystals are distributed at any size, whereas the histogram divides the entire size range into groups and displays the number of crystals in each group. The FBRM scans the crystal population, and the data is called the chord-length distribution (CLD), shown as a histogram. We generate an empirical model to relate the CLD and CSD histograms by use of experimental data to estimate the transformation matrix relating the two histograms. The CLD density function q(s) is deduced as a convolution form, as shown in eq 1.29

q L0(s) = qp(s , L0)n(L0) dL

∫0

q(s) =

where mcrystal is the mass that has been added to the saturated solution, ρ is the density of paracetamol crystals which equals 1.263 g/cm3, li is the geometric mean of the ith size range, and Vsolution is the volume of the solution, which is assumed to be the volume of the ethanol that was used to prepare the solution. From our inspection with an optical microscope, the crystals have an octahedral shape, and the aspect ratio of longest and second longest dimensions is around 2.0. Thus the shape factor kV is determined to be 1/3. (3) Fingerprint ai for the same size range can be obtained by

(1)

ai =

n(L0) is defined as the number density of crystals within an infinitely small size interval dL around L0, and n(L0) dL is the number of crystals at L0. Such crystals of size L0 generate a certain chord-length density qL0(s) through a transformation function qp(s,L0), which is the chord-

1 bj xi , j

(9)

where bj is the CLD measurement of the FBRM, which has unit [no. of chords] and xi,j is crystal concentration [no. of crystals/milliliter], the fingerprint ai carries the unit of [no. of chords milliliter/no. of 608

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

crystals]. After several additions, ai converges and its values are defined as the fingerprint for the ith size range. More details on the procedure can be found in our previous work.25 Model Implementation. Estimating the CSD histogram x from a known b requires the inversion of eq 5. In practice, the matrix A is often ill-conditioned. To deal with this problem, a regularization term is added to the least-squares minimization, as shown in eqs 10a−10e.

min A x − b x

s.t. xi ≥ 0, yi = li3

+ λf (y)

2

(10a)

i = 1, 2, ..., r

(10b)

xi , Li − Li − 1 r−2

f (y) =

i = 1, 2..., r

⎛ yi + 2 − 4yi + 1 + 3yi ⎞2 ⎟ , li + 2 − li ⎠ ⎝

∑⎜ i=0

li =

2

LiLi + 1 ,

(between 0.0 and 0.5 g solute/g solvent) and temperatures (between 0 and 70 °C). There are several approaches to quantify the solution concentration from the in situ spectra,32 including peak height/area regression, multiwavelength regression, principal component regression, and partial least-squares regression. In this study, we found that using peak heights is simple and sufficiently accurate. The AU at 1667 cm−1 is chosen as the characteristic peak of paracetamol and the AU at wavenumber 1048 cm−1 represents ethanol as shown in Figure 2 since such choices lead to good fitting between the peak height ratio and concentrations.

(10c)

y0 = 0

i = 1, 2, ..., r

(10d) (10e)

where r is the number of size ranges, Li and Li+1 are the boundaries of the ith size range, and li is the geometric mean of the boundary values for the average size in each size range. The first constraint (eq 10b) enforces non-negative values for crystal concentrations. The second constraint (eq 10d) defines a penalty term that suppresses the oscillation in the CSD estimate; y is defined as the volume- or massweighted number density. The penalty function f(y) is the sum of square of the forward difference approximation of the first-order derivative on the spatial direction. This function is multiplied by λ, which is the weight for the penalty term and chosen to be 1 × 10−10 (μm/mL)−2. In addition to information on the solids from the FBRM, the solution concentration is also investigated. As shown in Figure 1,

Figure 2. Infrared spectrum of paracetamol dissolved in ethanol.

The ratio h = AU(1667)/AU(1048) was correlated with temperature T (K−1) and molar percentage of paracetamol X as in the equation:

h = k1X2 + k 2X(T − 273.15) + k 3X + k4(T − 273.15) + k5 (12) The calibration experiments were carried out in a sealed and jacketed vessel. Solutions with known concentrations were cooled from high to low temperature, and the spectra were recorded. With known h, X, and T, linear fitting yields estimates of k1 = −8.781(−), k2 = 1.974 × 10−3(K−1), k3 = 3.648(−), k4 = −2.546 × 10−4 (K−1), and k5 = −2.819 × 10−4(−) with R2 = 0.93. The molar fraction can be calculated accordingly if h and T are obtained. The concentration estimated by IR, cIR, is then: cIR = Mw,paracetamolX/Mw,ethanol(1 − X). We also carried out an ad hoc correction against day-to-day variability, which are mainly caused by differences on the background spectrum and optical fiber curvature. The concentration readings from eq 12 were calibrated with known concentrations. In this study, in-run calibration is used by assuming that the concentration reaches the solubility value when the temperature is kept constant for a sufficiently long time. More specifically, the temperature is kept at 34 and 0 °C for about an hour in our experiment. Pure ethanol is also used for this calibration at X = 0 at temperatures from 0 to 20 °C. These three concentrations are used to linearly correct cIR against the day-to-day variability. With the calibration equation and the ad hoc correction, we can estimate solution concentration from the IR spectrum. The FBRM and ATR-FTIR measurements can be combined by the following mass balance equation

Figure 1. Solubility of paracetamol in ethanol. paracetamol is highly soluble in ethanol; its solubility was reported in several references.2,30,31 We suspect these values vary because of differences in methods and materials used in the experiments. Here, the lower bound of the solubility data was used to fit the third-order polynomial for the solubility cs [g solute/g solvent] at the temperature T (K): cs = 7.915 × 10−7T 3 − 6.439 × 10−4T 2 + 1.765 × 10−1T − 16.17

(11)

msolid + mliq = mtot

The infrared spectrum of the liquid phase is measured by attenuated total reflectance-Fourier transform infrared (ATR-FTIR) spectroscopy to estimate solute concentration. For paracetamol dissolved in ethanol, the spectrum depends on both temperature and concentration. We measured the IR spectra at various paracetamol concentrations

(13)

where msolid and mliquid are the mass of solute in solid and liquid phases, respectively, and mtot is the total amount when the solution is prepared. If mliquid and mtot are known, msolid can be calculated accordingly. The mass of the solid phase can be also obtained by 609

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

r

msolid(x est) ≈

∑ mcrystal,ixest,i i=1

histograms, the crystals were added to a 400 mL saturated solution at 24 °C. From the infrared spectrum, no change in solute concentration was observed on addition of the crystals. This indicated that the solution was well-equilibrated, and no change in crystal size occurred after addition to the solution.

(14)

where mcrystal,i is the mass per crystal in size range i. However, the CSD estimate xest by eq (10) may not always satisfy eq 13 due to estimation error. In particular, large crystals have a stronger influence on the mass estimate and thus any slight error at the large crystal size can cause significant differences in the evaluation of crystal mass. We introduce a modification by calculating a scaling coefficient: mtot − mliq x̂ = x est msolid(x est) (15)



RESULTS Fingerprint CLD Histograms. The fingerprint histograms of sieved paracetamol crystals were measured in saturated ethanolic solutions, displayed as number-based and mass-based versions in Figures 4 and 5, respectively.

which satisfies msolid(x̂) + mliq = mtot. We use x̂ as the estimate of CSD throughout this study. Since the solid phase information and solution concentration are connected via the mass balance, we need to synchronize the two measurements from their own sampling intervals. The CLD measurement interval is set at ten seconds, and the IR spectrum is taken every one minute. For each IR spectrum, four temporally closest CLDs are selected and their averaged CLD is calculated and defined as the CLD at the moment when the IR spectrum is recorded. Temperature values at each IR sampling time are obtained similarly. After the CSD estimates are obtained, a low pass filter, which is a moving average with the window width of 10, is applied to reduce the fluctuation of the estimates in the temporal direction.



EXPERIMENTAL SETUP

The experimental setup is shown in Figure 3. The ATR-FTIR is manufactured by Mettler Toledo, and it is commercially called ReactIR Figure 4. Number-based fingerprint histograms.

Figure 3. Experimental setup. iC10. The FBRM D600 is also from Mettler Toledo. All the other parts, including the temperature probe, stirring system, a 1 L glass crystallizer, and heating/cooling metal jacket, are assembled into the OptiMax from Mettler Toledo. The IR spectrum of the solution from 653 to 2998 cm−1 was measured every minute. The ReactIR, purged by compressed air, is equipped with a diamond probe on a AgX interface. The FBRM D600 was set in the fine mode, and the chord range from 1 to 1000 μm was divided logarithmically into 100 bins. The CLD histogram was recorded every ten seconds at 2 m/s scanning speed with 0 μm focal point, smoothed by an exponential filter. The temperature range was from 0 to 70 °C, and the stirring speed was set at 400 rpm. These instruments monitored the crystallization process and communicated with a computer via the iC software from Mettler-Toledo (iC FBRM 4.2.234, iC IR 4.3.27, and iControl 5.1.29). Paracetamol (Sigma-Aldrich, 99%) powder was dissolved in ethanol (Sigma-Aldrich, HPLC grade), some of which was recrystallized by batch cooling crystallization. The crystals are in form I (monoclinic), which were in the shape of octahedra by our micrograph observation. They were sieved into nine fractions (53−75, 75−106, 106−150, 150−200, 200−250, 250−300, 300−355, 355−425, and 425−500 μm) in a Ro-Tap RX-29 shaker. To measure the fingerprint CLD

Figure 5. Mass-based fingerprint histograms.

The number-based fingerprint CLD histograms are shown in Figure 4, on the basis of the number of crystals suspended in the unit volume of the solution as derived in eq 9. In the calibration of fingerprints, five or six additions are used and each addition has 0.2 to 2 g of crystals. The results are shown as thin lines, which exhibited spikes, especially in the 65.3−85.7 μm region, probably due to the fouling of the probe. This region usually included three to five points and was smoothed by linear interpolation according to the neighboring data. The 610

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

other slight oscillations were smoothed by a moving average filter using three data points to obtain the final fingerprints (thick lines). Figure 4 shows that the fingerprints become higher and wider when the crystal size increases because large crystals occupy more space and have greater chance to reflect the laser beam. Such an observation suggests that large crystals are more influential to the CLD than small crystals. The shapes and heights of the fingerprints indicate that this effect is very significant in our case. This property causes the ill-conditioning of the system, which requires the regularized least-squares method, as shown in eqs 10a−10e. Assuming the growth of crystals is the prevailing process in the crystallizer, the total chord count should increase, even though the number of crystals is approximately constant. Therefore, an increase in total chord count can result from growth and should not always be attributed to the birth of crystals. On the other hand, since kinetics processes such as ripening or agglomeration may happen only within the solid phase, we are also interested in the mass-based fingerprint histograms. The mass-based fingerprint histograms are directly calculated by dividing the number-based fingerprints by the mass of one crystal (ai/mcrystal,i). These fingerprint histograms have the opposite trend as shown in Figure 5; for a fixed mass of crystals, more chords are observed if that mass is made of smaller crystals. This is because, compared to the same mass of large crystals, small crystals can spread more pervasively, and therefore their chance of being detected is greater. If the crystals are agglomerating or ripening, for example, 1 g of crystals at 50 μm turns into 1 g of crystals at 100 μm due to agglomeration, we could expect that the total chord count would decrease. Model Validation. To confirm the linearity of the FBRM model, we experimentally tested to see if the total chord count of a CLD changes linearly with a change of the CSD. If eq 5 holds, then the following relationship is satisfied: p

pore opening

sample 2, x2

500 425 355 300 250 212 150 106 75 53 20 0

0.86 1.53 3.15 10.52 16.63 19.41 31.36 11.57 3.82 1.15 0.00 0.00

7.20 4.86 10.47 15.05 19.72 16.54 19.44 5.70 1.03 0.00 0.00 0.00

Figure 6. Total chord count after each addition. (16)

i=1

p i=1

Σi =p 1(Ax1,N)i

additions of sample 1 (P = 14). Six additions (Q = 6) of sample 2 also showed proportionality between mass of crystals and total chord count with the presence of sample 1. Notice that the slope of sample 1 is higher than that of sample 2, which agrees with the finding in the discussion of mass-based fingerprints that small crystals generate more chords per unit mass. The normalized CLD histogram measurement calculated from the final slurry (x20) was compared with the normalized simulation result of eq 5 in Figure 7. The consistency of these two normalized CLDs also validates the linear model. CSD Monitoring. An experiment of batch-cooling crystallization of paracetamol was monitored by both FBRM and ATR-FTIR, with the initial concentration at 0.385 g solute/g solvent. The cooling profiles were designed to include two cooling steps. The first one was cooling the clear solution to a temperature that created a high supersaturation and triggered primary nucleation. Then the temperature was held so that the crystals generated consumed the remaining supersaturation. The second cooling further decreased the temperature to the lowest operating temperature in order to induce further crystal growth in the absence of primary nucleation. After the temperature was held at the final value for about 100 min, two postrun actions were performed sequentially to confirm that the supersaturation had been completely

p

∑ bi = k1α ∑ (A x1,N )i + k 2β ∑ (A x2,N )i i=1

sample 1, x1

where the CLD starts to behave nonlinearly. Figure 6 shows that the linear trend is well-maintained through the 14

Here we use two crystal size distribution samples: sample 1 and sample 2. These two samples have their own mass-based CSD histograms, x1,N and x2,N. The mass of crystals in x1,N or x2,N is 1.0 g. If we change the CSD histogram by adding crystals either from samples 1 or 2, the CSD can be written as xy = kα1 x1,N + kβ2x2,N, where k1 and k2 are the masses of crystal from samples 1 and 2, and the superscript α = 1,2,...,P and β = 1,2,...,Q represent the time index of additions of samples 1 and 2, with the total number of additions γ = α + β. Hence xγ is the CSD histogram after the γth addition of crystals. Equation 16 can be reduced to p

mass percentage (%)

(μm)

p

∑ bi = ∑ (Ax)i i=1

Table 1. Compositions of Two CSD Histograms

i=1

(17)

Σi =p 1(Ax2,N)i

where and are constant. If we sequentially made P additions of sample 1 and then Q additions of sample 2, the total chord count should increase linearly, exhibiting two slopes, Σi p= 1(Ax1,N)i and Σi p= 1(Ax2,N)i. Samples 1 and 2 with CSD histograms shown in Table 1 were used to test the model in a saturated solution of 500 mL prepared at 24.2 °C. The mass-weighted mean sizes of crystals are 218 μm for sample 1 and 247 μm for sample 2. The first seven additions of sample 1 were all chosen to be less than two grams so that we could investigate if there exists a threshold 611

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

the concentration. At step (2) the concentration and solubility changed simultaneously and both increased by 0.002 g/g, which indicates there was little supersaturation before E. The supersaturation ratio (c/cs) during run 1 is shown in Figure 9. Note that from point B until the end of the

Figure 7. Comparison of model simulation and measurement of CLD of x20.

consumed: (1) adding paracetamol crystals to the slurry and (2) increasing temperature by 1 °C. The purpose of adding crystals is to eliminate any remaining supersaturation for the ad hoc concentration correction mentioned earlier. Furthermore, by comparing the concentration and solubility change after the slight increase of temperature, we can confirm our concentration prediction is accurate. Such a minor increase of temperature is assumed not to change the CSD significantly for the subsequent sieving test. The experimental procedure produced results referred as run I and run II. Since these two runs are similar, we discuss run I first and discuss run II only briefly. Run I: Temperature and Concentration. Temperature and concentration profiles for run I are shown in Figure 8. In the

Figure 9. Supersaturation profile of run I.

experiment, the supersaturation ratio was lower than 1.1. From the variation of supersaturation, we cannot infer whether secondary nucleation occurred, but this question will be revisited later. Run I: Tracking Key Parameters of CSD. The total number of crystals and mean size of the crystal population calculated from the CSD estimates are shown in Figure 10 (panels a and b). They are compared with the total chord count and mean length from the CLD histogram, which are commonly used to represent the number of crystals and size of crystals. Figure 10a shows that the total number of crystals obtained from the CSD estimates resulting from the developed model is more representative of the process than simply using the total chord count. For example, an increase in both the total chord count and the number of crystals indicate primary nucleation at A, but the subsequent total chord count shows a rapid increase at B that could be misinterpreted as nucleation. However, since the supersaturation ratio was close to 1 at B, as shown in Figure 9, the increase in chord counts is much more likely due to growth than to nucleation. This is an example of our analysis of fingerprint histograms in Figure 4: the total chord count highly increases with the size of the crystals (i.e., growth of crystals leads to an increase of the total chord count). Note that the estimated number of crystals was constant after B. When the process was approaching C, there was a slow increase in the number of crystals from 300 to 350 min, probably caused by secondary nucleation. Such a change cannot be seen from the total chord count. At D, neither of the total chord count nor the number of crystals responded to the addition of 2.0 g of crystals. Because these crystals were taken from the 100 g product of batch cooling crystallization operated similarly, it is reasonable that the 2% change in the population can hardly affect the CLD and the estimated CSD. At E, temperature was increased by 1.0 °C, which corresponds to a crystal mass decrease of only 1.5%, according to the solubility dependence on temperature. This minor dissolution of crystals was overestimated by the total

Figure 8. Temperature and concentration profiles of run I.

first cooling step from 70 to 34 °C, the concentration was constant as there was no crystallization. The concentration then started to drop drastically at point A, which is the evidence that primary nucleation occurred at about 34 °C, and the concentration rapidly decreased to the saturation value. Cooling was resumed at B and stopped at C when the temperature reached 0 °C. In the period from B to D, supersaturation increased modestly and was depleted when cooling stopped. In order to check the assumption of saturation at the end of the run, (1) 2.0 g of crystals were added at D and (2) the temperature was raised to 1.0 °C at E. Step (1) had no effect on 612

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

Figure 10. Comparison between CLD histogram measurement and estimated CSD histogram of run I: (a) total chord count and estimated number of crystals (black lines for guiding the eye) and (b) mean chord length and estimated mean size of crystals.

Figure 11. Estimated crystal number density of run I: (a) at A, B, C, and D. (b) Evolution of CSD estimates.

chord count, which dropped suddenly. On the other hand, there was no substantial change in the estimated number of crystals. For tracking the change of crystal size, the comparison in Figure 10b shows the capability of the mean size of the estimated CSD over the mean chord length. During the crystallization process, the crystal mean size was estimated from 130 μm at B to 175 μm at C, whereas the mean chord length varied only between 80 to 100 μm, which is not as large as that of the mean size. The CLD measurements prior to the red vertical line were considered as noise, since it was confirmed by visual inspection that no crystals had been formed before A. Run I: Validation of CSD Estimates. The estimated CSD density functions at A, B, C, and D are shown in Figure 11a, which are obtained by applying eq 10a−10e and then transforming the estimated CSD histogram to the number density. The evolution of the CSD densities is consistent with our understanding of the kinetics of the process. At A, crystals appeared due to primary nucleation. The CSD at B was obtained after the supersaturation accumulated in the first cooling step was consumed. From B to C, the density functions had approximately the same height, whereas the numbers of crystals at small sizes decreased and those at large sizes were greatly increased, which suggests the growth of crystals in the

second cooling step. The change from C to D, corresponding to the depletion of the remaining supersaturation after C as can be seen from Figure 9, also implied a slight growth of crystals. These changes from A to D are also shown in the threedimensional Figure 11b. In addition, we can see that crystals at the first bin (smallest size) started to increase slowly at C, which indicates secondary nucleation. The secondary nucleation may have been triggered because the supersaturation was accumulated to its maximum value at C. In the subsequent period to D, these newborn crystals became larger and moved to the second bin (75−106 μm), and secondary nucleation gradually ceased because of the depletion of supersaturation. Therefore, the number density of the smallest size dropped again. The estimated CSDs are demonstrated to be sufficiently accurate by comparing the final estimate with the sieving result, as shown in Figure 12. After the experimental run was completed, crystals were filtered, washed with toluene, and then dried in an oven. The mass histogram was obtained by sieving and recalculating to the number density. Only minor differences are seen between the estimated and measured population densities; the shape and magnitude of the density functions are remarkably similar. Moreover, the sieving result estimated the total number of crystals to be 2.72 × 104 per 613

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

temperature was increased by 1.0 °C at F′, the solute concentration increased slightly while the supersaturation ratio remained constant. From these observations, we conclude that supersaturation indeed reached 1.0, and the ATR-FTIR measurement and solubility model are accurate. Estimates of the crystal population density at each of the times identified in the previous paragraph were obtained by applying our model to in situ measurements of CLD data from run II shown in Figure 14. Crystals started to appear at A′,

Figure 12. Estimated crystal number density and sieving result of run I.

milliliter and mean size to be 150 μm while values estimated from our model were 2.42 × 104 crystals per milliliters and 141 μm, respectively. Run II: Different Cooling Rate and Postrun Analysis. The first cooling step from 70 to 34 °C in run II was reduced by 1/3 that of run I (0.2 °C/min vs 0.3 °C/min), while the initial concentration, instrument settings, and second-stage cooling rate were the same as in run I. The temperature and supersaturation profiles are shown in Figure 13. The rapid Figure 14. Development of CSD number density of run II.

which indicates primary nucleation. The crystal population developed to the CSD estimate of B′ at the end of the firststage cooling. Slight growth occurred in the temperature plateau from B′ to C′, where large crystals increased and small crystals decreased. The crystals grew significantly in the subsequent cooling stage, as shown in the evolution from C′ to E′. Table 2 lists key variables and observations associated with runs I and II, which were begun with identical conditions and operated identically, with the exception of the first-stage cooling rate. Run I had a wider metastable zone than run II (21 °C vs13 °C), and thus the supersaturation at primary nucleation in run I was much higher than in run II. Higher supersaturation at nucleation usually leads to higher production of a larger number of crystals, but in these two runs that seems not to have been the case; in other words, comparison of N1 (the number of crystals per unit volume at B in Figure 9 and C′ in Figure 13) for the two runs shows similar values. Interestingly, nucleation in run I occurred during the constant-temperature plateau, while in run II it occurred while the temperature was still decreasing. It is unclear whether these two factors (i.e., degree of supersaturation and cooling rate when nucleation occurred) played similar roles in producing similar numbers of crystals. Starting from the similar CSD, the second stages of the two runs were at the same cooling rate, and the supersaturation ratios for both followed similar patterns, which were shown in Figures 9 and 13. The maximum supersaturation ratios of 1.10 were achieved at the end of the second-stage cooling in the two runs. If secondary nucleation had been a major factor in either of the runs, the final population density function would be expected to be bimodal. Since that is not the case with either run, as confirmed by our model predictions and sieving at the end of run I, we conclude that secondary nucleation was a minor factor in determining the final crystal size distribution.

Figure 13. Temperature and supersaturation profiles of run II.

decrease in supersaturation at A′ (about 42 °C) is indicative of primary nucleation; the decrease continued until it became nearly constant. Cooling ended at B′, and the supersaturation again began to decrease. The second-stage cooling began at C′, and again the supersaturation began to increase until at D′ cooling was stopped. The effects of the postrun analysis are shown in the inset of Figure 13. They indicate the solution was saturated at around 350 min (i.e., at F′). To confirm this, we added 1.0 g of paracetamol powder (not crystals, as were added in run I). We chose to add powder in expectation of the greater surface area inducing more rapid consumption of any remaining supersaturation. (Microscopy confirmed the powder was of smaller size and had significantly greater surface area per unit mass.) Nevertheless, the supersaturation ratio did not drop significantly, even after this addition. Furthermore, when the 614

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

Table 2. Comparison between Runs I and II run

dT/dt (°C/min)

Tnuc (°C)

ΔT (°C)

Snuc

N1 (no./mL)

S2,max

Nend (no./mL)

I II

−0.3 −0.2

34 42

21 13

1.55 1.40

1.40 × 104 1.47 × 104

1.10 1.10

2.35 × 104 2.19 × 104

dT/dt is the cooling rate in first cooling stage. Tnuc is the temperature when primary nucleation occurred. ΔT = Tsat − Tnuc, where Tsat is the temperature at which the solution is saturated (S = 1). Snuc is the degree of supersaturation when primary nucleation occurred in the first cooling stage. N1 is the number of crystals at the end of the temperature plateau. S2,max is the maximum degree of supersaturation in the second cooling stage. Nend is the number of crystals at the end of the crystallization run (D for run I and E′ for run II).

Process Monitoring and Direct Control. The crystallization process can be monitored based on the CSD estimates, instead of the total chord count of CLD and mean chord length. Without quantified knowledge about the relationship between the CSD and the CLD, it may be ineffective to use the total chord count and the mean chord length as process indicators. Moreover, this technique can be used in a direct control; the number of crystals, the mean size, and other metrics of the CSD can be estimated to determine the cooling/ heating rate without a numerical model of crystallization. The robustness of the model may be influenced by several factors. The fingerprint model assumes the crystals have a fixed shape, but it neglects possibilities of shape changing processes such as polymorphic transformation, agglomeration, and breakage. The fingerprint model also assumes linearity. Although not observed in our study, at higher crystal concentrations, the linear assumption may not be justified. Nevertheless, as can be seen in Figure 12, the proposed model estimated the CSD sufficiently accurately, which may indicate that the above assumptions are valid in our case study. Parameter Estimation and Model-Based Control. With the FBRM and the ATR-FTIR, a full picture is drawn for a crystallization process, including temperature, concentration, and CSD, which are sufficient for estimating crystallization kinetic parameters by the population balance or its reduced model (method of moments), given that our estimates are consistent with sieving analysis. Once the parameters are obtained, we can predict the development of the CSD for a given cooling profile, and thus we can select the optimal profile to meet the requirement of CSD of the final product.33,34 Moreover, a model predictive controller can be programmed based on the parameters.

Figure 15 shows remarkable similarity between population density functions predicted for runs I and II.

Figure 15. Comparison between estimated crystal number density at D in run I and at E′ in run II.



DISCUSSION Advantages of the Model. The empirical model is simple to construct, and it can be inverted by applying regularized least-squares minimization. As we showed in our two example runs, the CSD of crystallization processes are well-monitored, not only agreeing with our understanding of the crystallization kinetics but also being consistent with the sieving results. This method is applicable to various crystallization processes, if the crystals have no high aspect ratio and the shape is fixed. For needle-shaped crystals, the method can still be valid but the issue would be finding a way of separating these crystals into different size ranges. Nature of the FBRM and the Empirical Model. We create the model based on a simple linear system, which directly relates CSDs to CLDs from experimental results. Therefore, the model depends on the technique of classifying the crystals with respect to their sizes. In this study, we used sieving to separate the crystals into nine size ranges and measured the fingerprint CLD histograms. The fingerprints also revealed the nature of the FBRM: the CLD measurement is dominated by the large crystals. There are two potential error sources of this methodology: (1) the model is a partial observer, which has a limited number of bins covering a limited size range and (2) the ill-conditioning of the model always brings difficulty in estimating small crystals, as we saw the overestimation in run II when powder was added. The former source is specific to the empirical model, which can be mediated by using more sieve trays. The latter one is a general issue for all kinds of crystals, which requires a tailored algorithm.



CONCLUSION



AUTHOR INFORMATION

In this article, an empirical model of FBRM measurements was used to map CLDs to CSDs. The model is intuitive and simple to build by using characteristic CLD histograms of different sizes of crystals as the fingerprints. The FBRM model was then applied to batch cooling crystallization, and our results demonstrated an ability to extract the CSD with the knowledge of the solution concentration. It successfully detected the onset of primary nucleation, secondary nucleation, and crystal growth. The framework developed here can provide the CSD information not only for batch cooling crystallization but also to general solid−liquid processes, where the fingerprint CLD histograms can be obtained. It is a very promising technique for process monitoring, parameter estimation, and control of crystallization.

Corresponding Author

*E-mail: [email protected]. 615

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616

Crystal Growth & Design

Article

Notes

(31) Hahnenkamp, I.; Graubner, G.; Gmehling, J. Int. J. Pharm. 2010, 388, 73−81. (32) Cornel, J.; Lindenberg, C.; Mazzotti, M. Ind. Eng. Chem. Res. 2008, 47, 4870−4882. (33) Nagy, Z. K.; Braatz, R. D. J. Process Control 2004, 14, 411−422. (34) Zhang, G. P.; Rohani, S. Chem. Eng. Sci. 2003, 58, 1887−1896.

The authors declare no competing financial interest.



ACKNOWLEDGMENTS The financial support from C. J. “Pete” Silas Chair, Georgia Research Alliance, and Consortium for Risk Evaluation with Stakeholder Participation (CRESP) is gratefully appreciated.



REFERENCES

(1) Barrett, P.; Glennon, B. Chem. Eng. Res. Des. 2002, 80, 799−805. (2) Worlitschek, J.; Mazzotti, M. Cryst. Growth Des. 2004, 4, 891− 903. (3) Schöll, J.; Lindenberg, C.; Vicum, L.; Mazzotti, M. Cryst. Growth Des. 2007, 7, 1653−1661. (4) Scholl, J.; Vicum, L.; Muller, M.; Mazzotti, M. Chem. Eng. Technol. 2006, 29, 257−264. (5) Hu, X. H.; Cunningham, J. C.; Winstead, D. Int. J. Pharm. 2008, 347, 54−61. (6) Tok, A.; Goh, X. P.; Ng, W.; Tan, R. AAPS PharmSciTech 2008, 9, 1083−1091. (7) Dowding, P. J.; Goodwin, J. W.; Vincent, B. Colloids Surf., A 2001, 192, 5−13. (8) Hukkanen, E. J.; Braatz, R. D. Sens. Actuators, B 2003, 96, 451− 459. (9) Monnier, O.; Klein, J. P.; Hoff, C.; Ratsimba, B. Part. Part. Syst. Charact. 1996, 13, 10−17. (10) Barrett, P.; Glennon, B. Part. Part. Syst. Charact. 1999, 16, 207− 211. (11) Heath, A. R.; Fawell, P. D.; Bahri, P. A.; Swift, J. D. Part. Part. Syst. Charact. 2002, 19, 84−95. (12) Tadayyon, A.; Rohani, S. Part. Part. Syst. Charact. 1998, 15, 127−135. (13) Fujiwara, M.; Chow, P. S.; Ma, D. L.; Braatz, R. D. Cryst. Growth Des. 2002, 2, 363−370. (14) Mitchell, N. A.; Frawley, P. J.; O’Ciardha, C. T. J. Cryst. Growth 2011, 321, 91−99. (15) Mitchell, N. A.; O’Ciardha, C. T.; Frawley, P. J. J. Cryst. Growth 2011, 328, 39−49. (16) Nagy, Z. K.; Fujiwara, M.; Woo, X. Y.; Braatz, R. D. Ind. Eng. Chem. Res. 2008, 47, 1245−1252. (17) Togkalidou, T.; Tung, H. H.; Sun, Y.; Andrews, A. T.; Braatz, R. D. Ind. Eng. Chem. Res. 2004, 43, 6168−6181. (18) Hermanto, M. W.; Braatz, R. D.; Chiu, M. S. AIChE J. 2011, 57, 1008−1019. (19) Saleemi, A. N.; Rielly, C. D.; Nagy, Z. K. Cryst. Growth Des. 2012, 12, 1792−1807. (20) Jiang, M.; Wong, M. H.; Zhu, Z. L.; Zhang, J. Q.; Zhou, L. F.; Wang, K.; Versypt, A. N. F.; Si, T.; Hasenberg, L. M.; Li, Y. E.; Braatz, R. D. Chem. Eng. Sci. 2012, 77, 2−9. (21) Li, M. Z.; Wilkinson, D. Chem. Eng. Sci. 2005, 60, 3251−3265. (22) Ruf, A.; Worlitschek, J.; Mazzotti, M. Part. Part. Syst. Charact. 2000, 17, 167−179. (23) Li, M. Z.; Wilkinson, D.; Patchigolla, K. Chem. Eng. Sci. 2005, 60, 4992−5003. (24) Abu Bakar, M. R.; Nagy, Z. K.; Rielly, C. D. Cryst. Growth Des. 2010, 10, 3892−3900. (25) Li, H.; Grover, M. A.; Kawajiri, Y.; Rousseau, R. W. Chem. Eng. Sci. 2013, 89, 142−151. (26) Yu, Z. Q.; Chow, P. S.; Tan, R. B. H. Org. Process Res. Dev. 2008, 12, 646−654. (27) Kail, N.; Briesen, H.; Marquardt, W. Powder Technol. 2008, 185, 211−222. (28) Kail, N.; Marquardt, W.; Briesen, H. Chem. Eng. Sci. 2009, 64, 984−1000. (29) Worlitschek, J.; Hocker, T.; Mazzotti, M. Part. Part. Syst. Charact. 2005, 22, 81−98. (30) Granberg, R. A.; Rasmuson, A. C. J. Chem. Eng. Data 1999, 44, 1391−1395. 616

dx.doi.org/10.1021/cg401484d | Cryst. Growth Des. 2014, 14, 607−616